Frog Crossing

Evolving a 15→8→4 neural net to hop up, left or right, or wait, across the traffic and along the logs.

A small neural network learns to get a frog across four lanes of traffic and a river to one of three home bays at the top. On the river it can only stand on the drifting logs, and a log carries it along, so it has to hop left and right as well as up. Reach a bay and it starts over from the bottom. The score is how far it safely gets: every row it reaches for the first time counts (a full crossing is worth twelve), so partial progress shows before it completes a crossing. The Genetic Algorithm (GA) evolves the network's weights between generations.

How it works

1. The network reads 15 inputs: whether the frog is on the road or a log and how long it can safely stay put; for the next row up, what kind of row it is and how many steps until a hop up, up-left or up-right would be safe; the same for the row after that; how far it is from the nearest home bay; and where it is across the screen and how far up. One hidden layer of 8 neurons feeds 4 outputs: hop up, left or right, or wait. The larger output wins.
2. Fast and slow lanes sit side by side, so a gap that is safe on one lane is not on the next. On the river, open water drowns the frog, a log that drifts too far carries it off the edge, and hopping into the top row anywhere but a home bay hits the hedge.
3. Every member of the population plays a run. Fitness rewards each new row and full crossings; getting hit, falling in or missing the bays ends the run.
4. The top performers are selected as parents. Their weights are recombined and mutated to seed the next generation.

What to watch

• Early generations hop into traffic or straight into the water.
• For several cycles the frog usually gets a crossing or two and then fails on the river, often by riding a log away from the bays.
• A strong policy waits for its gaps, steps along the logs to line up with a bay, and crosses again and again.